Trump's AI security framework excludes open models, leaves risks undefined

The Trump administration has finalized a voluntary AI testing framework designed to assess cybersecurity risks from advanced AI, but Axios reports it's both limited and strikingly vague. The framework only applies to closed-source AI models deemed to have 'state-of-the-art capabilities' and carrying 'national security risks'—yet it never defines those terms. It also explicitly excludes open models, meaning anyone can download and inspect their core components without federal review, and the framework cannot be used to restrict them after release. Companies including Anthropic, OpenAI, and Google attended a White House briefing on the finalized plan yesterday, but the administration isn't making the framework public.

This ambiguity is concerning for an initiative allegedly built to analyze precisely these threats. The framework establishes a 30-day grace period for government review of new models, but with no clear standards for what counts as 'frontier' or a 'national security risk,' compliance remains guesswork. While participation is voluntary, frontier labs like OpenAI and Anthropic have actively sought government guidance to avoid restrictions. Smaller AI providers that want to stay in the White House's good books could be especially hurt by the lack of defined criteria, potentially facing uneven enforcement or over-cautious self-regulation. The exclusion of open models also creates a regulatory blind spot, as widely available AI systems remain outside any formal oversight mechanism.